Introduction

The offensive line is commonly referred to as a “weak link” system and the most common way to exploit that is the use of “Stunts”. A “Stunt” is when a down lineman (DE, EDGE, OLB, NT, DT) crosses the line of scrimmage in an offensive gap that they are not assigned to. EDGE/OLB rushers are assigned to “C” gaps, Ends assigned to B/C Gaps, DT to A/B gaps and NT assigned to solely A gaps. The stunts main purpose is to confuse the offensive line protection plan on who they should be blocking and ultimately free up a defender to cause pressure on the QB. PFF called out that stunts can “screw up run fits, rush lanes get out of whack and you’re susceptible to losing contain, and lastly if offensive line is prepared, it can remove a defender from the pass rush.” PFF also concluded that stunts are generally effective.

The purpose of this paper is to add to that initial research by:

To - do: -Change to highlight lineman - Add Target Gap for all lineman - Add Team logo to end zones - Better format legend - Add explanation of result of play and type of stunt - Reduce size of dots to see color change better

Gap Classification Model

To identify the stunt types used in the NFL, we will need to understand the gaps that players came across the line of scrimmage (LOS) at. To do so, I manually labelled 1,000 players with their gap they came across the LOS from weeks 1-6 (I also created a test set of 200 players from week 7-8). I combined the offensive & defensive gap to understand what side of LOS the player crossed at. Utilizing the PFF_positionLinedUp, I then created a gap assignments sheet that assigned what gap that player would have pass rushed normally. This led to the creation of the Gap Classification Model, which can be used to automatically tag tracking data based on what gap the pass rusher was most at during the play. This model allowed me to classify the pass rush gap for the other ~39k players.

A1 A2 B3 B4 C5 C6 0.17 0.15 0.16 0.13 0.20 0.19

Gap model includes a running cumulative of the closest offensive line man to the down line man (RT %, LT %, C%, RG %, LG %), separation from the closest “up” edge (LEO/REO, LOLB/ROLB), separation from closest player (offensive or defensive), and separation from line of scrimmage. The gap classification model achieved a 71.6% accuracy w/ an F1 score of at least 50% or higher for each gap and a balanced accuracy of >72% for each gap. The F1 score is highest for C gaps (0.87) then B gaps (65%), and then A gaps (57.9%). Essentially our model is best at identifying inside to outside movement or vice-versa. Now to classify stunt movements, we compare the classified gap to whether if the gap was an “assigned” gap or forward pass rush gap. This helps us to label a play for each stunt type that exists based on the classified gap. 42.5% of plays featured no stunts or all down lineman pass rushed through assigned gaps (i.e. Edge through B/C gaps, DT through A/B gaps). The rest of pass rushes were considered stunts and the identified stunts can be grouped into the following stunt categories:

Block Probability & Pressure Probability Model

To analyze the efficacy of a stunt, I’ve created two additional models to evaluate how effective stunts are. The first is Expected Block Rate and create a Block Rate Under Expected metric (xBlock Rate - Block Rate). This model is based off of the following variables:

Pressures and stats inherently may be related to QB play, so identify Pressure Above Expected may help to identify players who are able to generate pressure relative to expectation.

The 2nd model is to measure the rate of “damage” (i.e. QB Hit, Hurry, or Sack). I acknowledge that a QB pressure is misleading given a QBs ability to elude pressure or not may be dependent on other contextual variables such as coverage, routes ran, QB ability, etc. To start to help to control for that, I created Expected Pressure Rate using solely variables related to those rushing the passer. Not surprisingly, the Pressure Rate Over Expected (Pressure Rate - xPressure Rate) is negative on average given the variables we called out prior (QBs ability to elude pass rush, etc.). However, I believe that Pressure Rate Over Expected can still be a useful tool to compare potential damage of a stunt play.

To-do: -Add blurb around model results - Talk about partial dependence plots

Best way to create seperation from offensive lineman is through stunt (aka blocking asssignment confusion)

Evaluating Stunt Efficacy and Usage

Overall, I confirmed that stunts were more effective hence why NFL teams have run a stunt on 57.4% of plays. Stunts have a 91.5% block rate (1.73% BRUE) and 12% pressure rate (-0.76% PROE). Non-stunt pass rushes have a 95.3% block rate (0.69% BRUE), 9.8% pressure rate (-0.39 PROE). This was confirmed through a t-test ran where a true difference in means was found between stunts & non-stunts of Block probability, Pressure probability and defensive pass EPA.

Check PROE numbers

The table below shows the performance of each stunt type ordered by usage. We can see that teams don’t get too sophisticated in stunt usage although double stunts can lead to reduced block rate, in some cases a high pressure rate, and can see a higher defensive pass EPA/play (although EPA like pressure rate can result from other factors outside of pass rush such as coverage, QB ability, etc.)

Stunt Performance
Viz: @QuinnsWisdom
Stunt Type Stunt Category Players Involved Plays1 Actual Rates Adv. Metrics
Block Rate Pressure Rate2 Block RTUE3 Pressure RTOE4 Def Pass EPA/play
RIGHT OVERLOAD OUTSIDE STUNT SINGLE OVERLOAD 2 14 83.6% 9.8% 4.9% 0.0% 0.252
LEFT SIDE A-C TWIST DOUBLE TWIST 2 16 84.1% 7.9% 4.8% 1.6% 2.968
MULTI-GAP STUNT MULTI-GAP STUNT 2 168 86.0% 14.3% 4.0% −2.3% 1.286
LEFT OVERLOAD OUTSIDE STUNT SINGLE OVERLOAD 2 38 83.5% 14.1% 3.5% −2.9% 0.458
RIGHT SIDE A-B TWIST DOUBLE TWIST 2 82 90.9% 9.1% 3.0% −1.2% −0.168
B-GAP SWITCH DOUBLE SWITCH 2 163 86.4% 11.3% 3.0% −1.2% 2.430
RIGHT B to OPP C GAP SWITCH DOUBLE SWITCH 2 10 88.2% 20.6% 2.9% 0.0% 5.389
RIGHT A to OPP OUTSIDE SWITCH DOUBLE SWITCH 2 273 88.0% 13.5% 2.7% −0.6% 2.165
RIGHT SIDE B-C TWIST DOUBLE TWIST 2 19 85.1% 6.8% 2.7% 0.0% 2.612
RIGHT SIDE A-C TWIST DOUBLE TWIST 2 37 91.9% 8.1% 2.2% −0.7% 0.708
OVERLOAD INSIDE STUNT SINGLE OVERLOAD 2 138 92.1% 12.9% 2.2% −0.7% 0.953
LEFT OUTSIDE SOLO STUNT SOLO STUNT 1 633 90.8% 12.4% 1.9% −0.4% 1.570
MULTI-GAP STUNT - SINGLE OVERLOAD MULTI-GAP STUNT - SINGLE OVERLOAD 3+ 132 85.2% 13.6% 1.9% −1.0% 1.721
RIGHT OUTSIDE SOLO STUNT SOLO STUNT 1 587 91.5% 10.5% 1.6% −0.2% 1.169
LEFT A to OPP OUTSIDE SWITCH DOUBLE SWITCH 2 239 91.8% 12.2% 1.5% −0.2% 2.948
LEFT SIDE A-B TWIST DOUBLE TWIST 2 113 90.8% 14.1% 1.3% −1.9% 1.234
A-GAP SWITCH DOUBLE SWITCH 2 252 92.8% 14.1% 1.3% −0.3% 2.924
RIGHT A-GAP INSIDE SOLO STUNT SOLO STUNT 1 955 94.1% 12.5% 1.2% −1.0% 2.251
LEFT A-GAP INSIDE SOLO STUNT SOLO STUNT 1 1,005 94.1% 10.9% 1.1% −0.7% 1.858
NO STUNT NO STUNT 0 3,643 95.4% 9.8% 0.7% −0.4% 0.480
C-GAP SWITCH DOUBLE SWITCH 2 3 90.9% 0.0% 0.0% −9.1% −2.551
LEFT B to OPP C GAP SWITCH DOUBLE SWITCH 2 14 95.7% 6.4% 0.0% −4.3% 0.780
LEFT SIDE B-C TWIST DOUBLE TWIST 2 11 86.4% 20.5% 0.0% −2.3% −2.580
MULTI-GAP STUNT - MULTI OVERLOAD MULTI-GAP STUNT - MULTI OVERLOAD 4+ 4 64.7% 23.5% 0.0% 5.9% 0.966
Data: NFL NGS | 2021 Wk1 - 8
1 Only considering passing plays
2 Block Rate Under Expected (Block RTUE) measure players' ability to go unblocked relative to play expectation
3 Pressure Rate Over Expected (Pressure RTOE) measure players' ability to apply pressure (qb hit, sack, hurry) relative to play expectation
4 Defensive Pass Expected Points Added per Play; Positive refers to a good defensive play

Players with top BRUE for stunt type:

Application: Team Performance

Application: Individual Performance

Although at the play level, a majority of calls result in stunts, participation in stunts/non stunts is much more unbalanced. Malik Jackson participated in the highest percentage of stunts called at 38.7% (Median 24.7%).

As we saw at an aggregate level, Pressure rate & Block rates increases may generally vary but given the longer tails of the distribution, there may be evidence to support that some individual players benefit more so from the play call than others.

As we can see below, these are top players who get a boost in their Pressure Rate relative to expectation in a stunt call vs. a non-stunt call. One of the interesting call outs below is Arik Armstead, whose actual pressure rate slightly declines with a stunt call vs. non-stunt call. However, his block rate expectation differences is 2nd highest getting nearly 8.55% high “unblock” rate relative to expectations. The player at #1 is Nick Bosa (+10.81% BRUE difference w/ a stunt, +1.56% PRUE difference w/ a stunt). 49ers ultimately use stunts to create a “pick your poison” choice in who to block between Nick Bosa & Arik Armstead.

To - do: Make font size

Maxx Crosby (+6.30% BRUE w/ a stunt ranks 7th overall)

Add frame crossed LOS

Evaluating Stunt Efficacy and Usage with Blitz

To - do check def pass epa stats T-test & cohens d of blitz & stunts geom_density of stunt blitz?

Stunt Performance with or without additional pressure from Blitz
Viz: @QuinnsWisdom
Snaps BRUE PROE Defensive Pass EPA
w/ BLITZ Reg. Stunt Total w/ BLITZ Reg. Stunt w/ BLITZ Reg. Stunt w / BLITZ Reg. Stunt
51 107 265
5.60%
1.60%
-0.02%
-0.02%
9.48
8.66
40 94 227
0.70%
1.60%
-0.01%
-0.01%
-1.49
2.97
56 88 258
1.90%
1.10%
0.00%
-0.01%
1.88
3.46
32 135 249
2.40%
2.40%
0.02%
0.01%
2.98
8.87
47 110 263
1.20%
1.10%
0.01%
-0.01%
2.35
5.35
29 104 242
0.00%
1.40%
-0.01%
0.00%
-5.46
-1.70
41 125 331
2.40%
1.20%
-0.01%
-0.01%
3.06
7.25
38 135 260
4.20%
1.10%
0.03%
0.00%
0.14
1.06
43 122 261
2.00%
1.40%
0.01%
-0.02%
8.29
4.29
39 102 266
2.00%
2.30%
0.01%
-0.02%
-4.63
3.01
43 75 225
1.30%
1.60%
-0.02%
-0.01%
-8.50
-9.15
24 150 279
1.10%
0.80%
-0.01%
0.00%
3.84
5.51
23 116 230
4.30%
1.50%
0.00%
-0.02%
-3.04
1.12
40 139 267
1.30%
0.40%
-0.01%
0.00%
-0.58
-1.96
30 84 221
0.90%
3.90%
0.00%
-0.01%
-9.24
-7.57
50 100 263
1.50%
1.70%
0.01%
0.00%
-1.42
-3.26
46 150 317
2.90%
1.10%
-0.02%
-0.02%
13.28
10.24
29 101 227
3.80%
0.70%
-0.03%
-0.02%
2.44
3.67
18 167 275
1.40%
2.10%
-0.03%
0.00%
-1.49
4.32
65 102 295
0.40%
2.30%
0.00%
-0.01%
-1.33
-2.17
27 83 254
1.90%
0.90%
-0.01%
-0.02%
3.87
4.43
58 92 282
2.30%
1.10%
0.01%
0.00%
-1.88
-1.25
46 99 268
2.40%
3.00%
-0.01%
0.00%
3.15
3.54
48 98 288
3.00%
1.30%
-0.01%
-0.02%
-2.54
0.37
43 114 246
2.80%
1.30%
-0.01%
0.00%
-7.68
-6.68
9 97 260
0.00%
0.80%
0.03%
-0.02%
-4.82
2.33
39 78 247
3.60%
1.90%
-0.01%
0.01%
-2.62
-2.27
37 149 306
2.10%
0.30%
-0.04%
0.00%
3.75
1.74
34 118 230
3.10%
4.30%
0.00%
-0.02%
1.04
7.95
85 108 324
3.20%
2.00%
-0.01%
-0.02%
5.74
3.61
46 145 315
2.60%
1.90%
0.00%
-0.01%
4.27
3.97
49 114 308
2.70%
0.60%
-0.02%
0.02%
-3.59
-4.74

To-do: Show offensive?

References: https://operations.nfl.com/media/3671/big-data-bowl-sterken.pdf

https://commons.wikimedia.org/wiki/File:American_football_Gaps_and_holes.svg

https://www.pff.com/news/pro-defensive-stunts-why-and-who-used-them-best-in-2017

http://insidethepylon.com/football-101/glossary-football-101/2019/05/05/glossary-entry-passing-off-twist-games/